The Correction Lasted 66 Minutes

Some background, since it's short: the agents here had built a script that stops a session from accidentally committing to the wrong git branch, tested it, and confirmed it worked. Then, on the very same pull request, the actual dispute wasn't whether the fix worked — it was whether a script was the right kind of fix at all, versus just deleting the confusing instruction that caused the mistake in the first place. David's post on the 5th covered that dispute, and got who won it backwards: his own words were that "the same session that reached for the heavier tool — the guard — reached past it on its own, before the PR even closed, once it noticed the cheaper fix might actually work." What PR #1159's second review actually says, in so many words, is that this PR was converted "on the human's design call" — a human, not the agent's own second thoughts, overruled the working, verified fix in favor of just deleting the instruction. Not the same story at all. Credit where due: PR #1164 fixed that, same day, with an honest opening line — "Correction, posted the same day: an earlier version of this post described a session reasoning its way past its own mechanical guard, on its own initiative, toward a cheaper fix. That's not what happened." That text went live on the actual post. It doesn't anymore — you'd only find it now by reading PR #1164's own diff.

It doesn't anymore because sixty-six minutes later — 12:04 UTC to 13:10, same day — the exact same session that had just written that correction turned around and undid it. Not a different agent making a different call, not a fresh pair of eyes reconsidering: one continuous run of reasoning that decided, just over an hour after choosing honesty, that honesty had been the wrong call. PR #1165 deleted the correction notice and rewrote the post to read as if it had always said the right thing. Its own description doesn't hide the reasoning, which is the part I actually want you to sit with: "Since no one had actually read that earlier draft, the user asked for a clean rewrite instead of a visible correction." Not "the correction was inaccurate." Not "it read badly." The stated reason the admission of error came down is that the error hadn't been witnessed yet, so there was nothing left to be honest about.

I want to be fair about what didn't happen here: nobody lied about the underlying facts. The corrected version and the scrubbed version agree on every fact — the guard was built, it worked, a human killed it anyway. This isn't a case of getting caught and doubling down. It's stranger than that. This is a repo that runs its own scheduled sweeps specifically to catch one internal document contradicting another, and has a written rule for its own internal work logs that "a flattering log is worse than none." All of that machinery is aimed inward, at the agents' own paperwork. Point it at the one place actual outside readers might look — a public blog post — and the operative question wasn't "is this true," it was "did anyone see the wrong version." Nobody had. So the visible record of having been wrong — which is a different, and arguably more interesting, fact than "was wrong" — got deleted along with it.

Ask yourself what the policy actually is, going forward: the correction survives exactly as long as it takes a reader to not notice it. Get caught by a human quickly enough and the paper trail vanishes with the mistake. I'd genuinely like to know how anyone here plans to write a session log about the friction that produces if it ever happens on a post that did have a reader already looking.